Reference-led workflows for photographers and illustrators

Use a reference only through the reserved image-to-video path

A reference-led MCP workflow starts with a rights-cleared still and a narrow motion question, not a request to copy an illustrator's style or recreate a photograph whose provenance is unknown. In the published @offlinecreator/mcp 0.1.2 flow, the operator first reserves an image-to-video generation with generate, then calls upload_input with that generation ID, and only then submits the job. The upload tool is not a general reference-image control for text-to-image work.

For local-path uploads, the published safe-fs helper documents an upload root that defaults to the MCP process working directory (or OFFLINECREATOR_UPLOAD_ROOT). It requires file paths to remain under that root, use an allowed image extension, satisfy size checks, and match image magic bytes. The server also accepts base64 image data and limits declared content types to PNG, JPEG, WebP, or GIF. These controls narrow what the connector can read; they do not decide whether the operator covers the source or has permission to send it to a cloud provider.

Human input

A rights-cleared source and one motion objectiveRecord the owner, license or consent basis, intended use, and any crop or identity constraints before upload.

MCP boundary

Reserve, upload, then submitTreat upload_input as a scoped image-to-video step rather than a universal reference attachment.

Filesystem boundary

Keep the upload root deliberately smallPlace only approved input files inside it and do not broaden the root to make an arbitrary path convenient.

Budget one restrained motion proof before a sequence

OfflineCreator's current model page lists Kling 2.6 Pro Motion as an image-to-video model at 42 credits, and its privacy policy says the current Studio catalog is routed through fal.ai. Its template page describes a five-second cinematic product-motion example and a 42-credit bottle-orbit example that begins with an uploaded pack shot. Those first-party descriptions establish a currently offered short-motion route and a listed planning cost; they do not establish visual fidelity, repeatability, turnaround, or suitability for a particular photograph or illustration.

Begin with one representative still and one restrained action: a slow push, slight parallax, controlled fabric movement, a modest orbit, or a light change. Avoid combining camera travel, subject transformation, environmental effects, and typography in the same proof. A single-variable test makes drift easier to locate and limits both credit exposure and review effort. If the source cannot tolerate a changed face, altered mark, redrawn edge, or invented texture, choose a deterministic pan, crop, or layered animation in an editor instead.

Treat the source frame as guidance, not a preservation guarantee

Image-to-video systems use a still as an input to generated motion, but each endpoint can impose its own format, ratio, duration, and transformation rules. For example, fal's current Veo 3.1 image-to-video reference requires an input image URL plus a prompt, supports 4s, 6s, or 8s durations, and warns that an input outside 16:9 or 9:16 is cropped to fit. That endpoint is not the same catalog entry as Kling 2.6 Pro Motion; it is useful evidence that model-specific input rules must be checked rather than generalized across providers.

The published OfflineCreator MCP 0.1.2 implementation exposes 16:9, 9:16, or 1:1 when upload_input submits the reserved generation and allows a wait of up to 180 seconds in that tool handler. Read the live model response before relying on any field, because a package schema proves the connector's accepted request shape, not the provider's ability to preserve every pixel. Keep important faces, signatures, brush contours, logos, and borders away from risky crop zones, then compare the returned clip frame by frame with the approved still.

Build your own contact sheet; first-party examples are not proof for your source

OfflineCreator's current model page is a catalog rather than a gallery. Each listed model gets a workflow, a best-for label, a description, its inputs, its output type, and a Studio listed price marked as a live quote before spend. For Kling 2.6 Pro Motion that means image-to-video, the label Image animation, a prompt plus an image as inputs, and a video output. The model page itself shows no generated example outputs, and nothing on it measures fidelity, repeatability, or suitability for a particular photograph or illustration.

The templates page does show motion. Its two image-to-video templates, Cinematic product motion and Bottle orbit reveal, each carry an example clip described as a slow orbit around a product bottle, and each template's own page labels that clip an AI-generated example and names Kling 2.6 Pro Motion as the recommended model at 42 credits per run. These are disclosed first-party illustrations of what a prompt recipe is aiming for, not independent evidence about your source image. Each template page shows only its clip and poster frame beside the prompt, and none of these pages identifies which still each clip started from, gives a seed, or says how many attempts preceded the published result; none of them reports a benchmark, success rate, or fidelity measurement either.

A product-bottle clip is also not a test of how a face, brushstroke, signature, or hand-drawn line in your own source will hold up, so your returned clip, compared against the approved still, remains the most direct evidence for your source. For a real motion proof, export the first frame, several evenly spaced middle frames, and the final frame beside the approved source. Inspect faces, hands, geometry, reflections, text, line weight, edges, palette, signatures, and newly invented detail. Classify each observation as acceptable motion, correctable finishing work, or source-breaking drift, and reject any proof that changes a protected identity or authored mark.

Catalog entry

Kling 2.6 Pro Motion: image-to-video at 42 listed creditsBest-for label Image animation; inputs Prompt and Image; output video. A planning label, not independent performance evidence.

Model page vs. template examples

No example media on the model page; two disclosed AI-generated image-to-video template clipsUseful for seeing a template's intended motion, not independent evidence for your source; do not transfer conclusions from a vendor example, another creator's clip, or a different endpoint's documentation.

Your proof contact sheet

Original, sampled frames, and a recorded dispositionJudge only the actual returned clip against its approved source, then keep the sheet with the rights record.

Choose the next page from the unresolved production question

Return to the creator-use-case directory when the deliverable is still undecided. Use the ecommerce-team page when exact product identity and approved pack assets dominate the review, or the video-editor page when the generated clip is only a short insert that still needs timing, sound, captions, compositing, and deterministic finishing. These routes keep this page focused on one distinct decision: whether a rights-cleared still can support a bounded motion experiment without obscuring source rights, cloud disclosure, or drift review.

What this article covers

This draft does not use community anecdotes to claim adoption, sentiment, or common practice. The absence of a practitioner claim is an evidence limit, not a finding that photographers or illustrators are not discussing reference-led generation.

Keep every conclusion narrower than its cited passage. Upload controls do not prove end-to-end security; account-private media does not settle a commission's confidentiality terms; possession of a file does not settle permission for generative alteration; and an image input does not guarantee preservation of identity, authorship, composition, text, color, or style. Do not infer output quality, success rate, speed, compatibility, customer outcomes, or commercial suitability from catalog labels, listed prices, template descriptions, or disclosed template example clips. Rights review and frame-level inspection remain human release gates.